A company name does not identify your work sample. Confirm the environment, permitted tools, expected artifact and scoring. Choose practice that reproduces those constraints.
Problem
A fictional recruiter packet confirms two rounds:
| Round | Environment/time | Tool rule |
|---|---|---|
| Implementation | Existing codebase, 45 minutes | One named embedded assistant allowed |
| Oral math | Spoken derivation | No AI |
Choose the drills. Does the first permission authorize an outside assistant or tools in the second round?
Picture: Route by Evidence, Not a Logo
Confirmed constraints choose a drill; unknown tool permission prompts a question
Derivation
Does knowing the time budget grant unknown tool permission?
Work the Packet
- Practice code navigation, small edits, tests and review within 45 minutes using only the authorized assistant.
- Practice closed-book derivations and changed-assumption follow-ups for oral math.
- Do not transfer permission between tools or rounds. Clarify any conflict with public policy before assessment.
The preparation decision tree
Ask the recruiter six questions:
- What are the exact round names and durations?
- Is implementation blank-file, codebase, notebook, take-home, or trace-based?
- Which AI tools are allowed in each round?
- Is there a presentation, research discussion, or artifact review?
- Which domain and team will technical depth target?
- What does each round produce or score?
Then map format to drill:
- unfamiliar codebase plus AI: agentic codebase lab;
- LLM training debugging: broken training lab;
- research notebook or API: black-box lab;
- project presentation: presentation protocol;
- performance trace: accelerator lab;
- ML primitives: implementation set;
- oral math: math lab;
- values or mission: values packet.
Change 1: One Round Is Silent
New task. A second implementation round gives no tool rule.
Picture and change. Permission belongs to a round/tool pair, not the whole loop. Ask before using tools.
Check. The first round’s allowance does not propagate.
Use when. Instructions differ or are incomplete.
Change 2: A Report Describes a Pilot
New task. A dated account describes a new format.
Picture and change. Keep its role/date scope; use it to ask about your round, not assert a stable process.
Check. An AI-use questionnaire, coding assistant, reported pilot and AI grading are four different claims.
Use when. Official detail is sparse.
What not to do
Do not route by brand, practice every lab or memorize alleged leaked questions. Jobs describe work scope, not exact prompts. Fellows or pilot evidence stays within that population.
Transfer Check
Which format-specific drill follows from knowing only the employer name?
Reasoned Answer
None yet. Confirm role/round constraints; general foundations are only a provisional baseline.
Evidence quality matters
Separate five evidence classes:
- Official general policy: organization-wide context, not every round’s detail.
- Current role/round instructions: scoped constraints; clarify conflicts with an authorized contact.
- Method-described reports: inspect recency and population.
- First-person accounts: apply only to that candidate’s role/date.
- Unverified anecdotes: questions to ask, not facts or permission.
Current format map
These primary pages were checked on 2026-10-24, replacing the inconsistent July/August verification claim. This is a check date, not the pages’ publication date or a guarantee about your round.
| Lab | Evidence class | Checked primary-page statement | Unknown for your round |
|---|---|---|---|
| OpenAI | Official general guide | Formats and AI rules vary by interview; prep materials specify allowances | Exact format, tools and duration |
| Anthropic | Official candidate policy | No AI during take-homes/live interviews unless explicitly allowed | Exceptions in your instructions |
| Google DeepMind | Official typical overview | Role-specific steps; overview includes two or three skills calls | Exact work sample and tool rule |
| Meta | Official general process | Select roles include an authorized assistant within CoderPad | Role coverage, models and design rules; FAQ answers were not extracted |
| xAI | Official general process | Technical application review, screening and technical-depth sessions | Exact ML round list and submission length |
Anthropic’s page displays 2025-07-10 as its update date. Publication dates for the other extracted pages were not established. Recheck before scheduling/practice decisions and when instructions change. Never treat a cached summary as permission.
Source Record
Record company, team/role, round, format, tools, artifact, duration, scoring, source URL, evidence class, checked-at date, scope and caveat. Unknown is valid; resolve consequential gaps before using tools.
OpenAI
The interviewing.io report was previously described as June 2026 reporting, including a pilot. That historical description was not newly verified here.
Anthropic
Careers, the performance write-up, original challenge and Fellows posting are distinct reference scopes, not a universal assessment.
Google DeepMind
The Research Engineer synthesis is a secondary report, not official instructions.
Meta
Do not infer outside-tool permission from an embedded assistant.
xAI
Confirm application constraints directly; the checked overview does not establish a universal 100-word limit.
See the source-labeled matrix; inspect each record’s date and caveat separately.
Related: frontier-lab process matrix, role-specific paths, and executable labs.